Do you have a question? Want to learn more about our products and solutions, the latest career opportunities, or our events? We're here to help. Get in touch with us.
Traditional contact centre planning models are built on predictable human behaviour, and the emergence of AI-literate customers and machine callers is breaking those assumptions in ways that current metrics and forecasting approaches aren't equipped to detect.
Four operational problems are becoming imminent: synchronised demand spikes from machine callers, identity verification processes that can't distinguish authorised AI from malicious AI, fallback cascades that add effort rather than removing it, and a knowledge gap that inverts the traditional agent-customer dynamic.
Existing KPIs like average handle time, deflection rate and service level are providing a misleading picture of performance, and contact centres need new measurement frameworks that account for the changing composition of demand.
The operational models underpinning contact centre planning have been remarkably stable for decades. We spoke to Brendon Edwards, Associate Director of Workforce Optimisation at Datacom, about what happens when the foundational assumptions those models rely on start to break down, and how workforce management needs to evolve in response.
Every contact centre is built on patterns. Predictable arrival curves that reflect human behaviour. Erlang models that rely on the natural variance of when people decide to pick up the phone. Training frameworks designed around agents holding the knowledge advantage. Quality assurance that measures whether an agent followed the script.
The morning arrival curve in any contact centre is a representation of how humans move through their day. School drop-offs, commutes, lunch breaks, procrastination. For decades, that variance has been stable enough to forecast around, but the question now is what happens when the callers stop behaving like humans.
The shift toward AI-literate customers, and increasingly toward AI agents acting on behalf of customers, is creating operational challenges that traditional workforce management approaches weren't designed to handle. Four of these are imminent, and current measurement frameworks will struggle to surface them until something breaks.
Human callers create natural variance. They call at different times because they have different routines, different constraints, different levels of urgency. That variance is what makes demand forecastable.
Machine callers operate with precision. If a thousand customers instruct their personal AI assistant to "call my energy provider when they open and sort out that bill," those calls arrive all at once, at 8:01am – rather than gradually across the morning.
Even if machine callers represent only 5% of daily volume arriving in the first 15 minutes, that concentration has consequences. It either blows your service level in the opening intervals or requires overstaffing that destroys efficiency for the remainder of the day. Erlang calculations aren't built for demand that arrives with machine precision rather than human variance.
There's no clean operational response to this yet. We're beginning to think about whether separate service levels for machine-originated and human-originated contacts make sense, and what the composition of demand might look like as personal AI assistants become more capable and more widespread.
Alongside synchronised demand, there's the latent demand problem. Reducing friction for customers doesn't reduce volume in the way many business cases predict. When it becomes effortless to pursue a complaint or query, people who previously would have let an issue go now have the tools to act on it. The bar of demand doesn't get shorter. It changes shape.
Identity and verification processes were built to confirm that a human caller is who they claim to be. They work on the basis that someone who can provide their date of birth, address and account details is likely the account holder.
A machine caller with access to a customer's credentials can present them flawlessly. The verification passes, but a critical question remains unanswered: did the customer authorise this interaction? Is the machine acting in their interest?
If a legitimate AI assistant can pass your ID&V checks, a malicious one can do the same. Research estimates AI-powered fraud will cost businesses more than $10 billion annually by 2027. The cost to build an AI agent capable of passing a standard identity check is roughly $20 per month. We are using 20th-century verification to guard against 21st-century impersonation.
Two questions emerge from this that every organisation will need to answer. The first is about policy: do you accept machine callers for all interactions, a subset of interaction types, or not at all? The second is about detection: how do you differentiate between a human and a machine caller, and between a machine acting on legitimate authority versus one that isn't?
The fallback cascade is the scenario that concerns me most from a workforce planning perspective. It describes what happens when a machine caller engages with a human agent, the interaction reaches a complexity threshold the machine can't navigate, and control is handed to the customer themselves.
The nature of the interaction changes mid-conversation. Additional handle time is introduced, but the harder question is what information the machine passed to the customer before the handoff. Did it provide a script? Additional context to support a negotiation? Or was it a cold transfer that leaves the agent restarting from scratch with a frustrated customer?
Gartner research indicates customer effort increases by 40% when a digital-to-human handoff requires repetition, and churn propensity increases with each additional handoff. From a WFM perspective, forecasting for fallback cascades is difficult. Do we forecast by personal assistant type? By some distribution of AI platform market share? The variables are still emerging and we don't have established models for them yet.
The knowledge divide describes the scenario where a customer, whether human or machine-assisted, has access to a richer and more current knowledge base than the agent. Competitor pricing, ombudsman decisions, current legislation, media coverage, review aggregators, all synthesised and available in seconds.
Meanwhile, agents have access to the internal knowledge base. In many organisations, those bases aren't kept as current as they should be. The conversation dynamic shifts from one where the agent informs the customer to one where the customer arrives pre-informed and ready to dispute or negotiate from the opening line.
Historically, agent knowledge has evolved through several phases. Memory recall and regurgitation of memorised policies. Knowledge base lookup and relay. Behavioural training that emphasises listening, searching and responding. The next evolution involves augmenting agents with tools that give them access to a knowledge space at least as rich as what the customer already has. The underlying question is how much external access those augmented tools should have, whether they stay confined to internal knowledge or draw from the same open sources customers are using.
These four problems compound into a broader challenge: most of the metrics contact centres rely on are providing a misleading view of performance because they were designed for a world where both sides of the interaction were human.
Average handle time is increasing across the industry, but the cause is compositional rather than performance related. As simpler work shifts to automation, the remaining human workload concentrates around complexity and emotion. A complexity-adjusted AHT provides a more honest picture of whether agents are performing well against the work they're receiving.
Deflection rate is, in my view, the most dangerous metric in the industry. Deflecting a contact tells you nothing about whether the issue was resolved. If a portion of deflected volume returns 48 hours later as a more complex human contact, the original deflection wasn't a resolution. It was a deferral. Net resolution rate, which measures whether the issue reached a conclusion regardless of channel, is a more useful measure of whether automation is working.
Service level can look artificially healthy when AI handles a large percentage of contacts instantly, while the human queue behind it is in crisis. Reporting human service levels separately from machine-to-machine interactions provides visibility into what the customer is experiencing, rather than what the blended average suggests.
Abandon rate is being distorted by machine callers that hit the IVR, don't get through, disconnect, and retry systematically. In the data these look like abandons, but they're failure-to-serve retries. An attempt-to-resolution ratio captures the full journey more accurately.
No organisation has all of this figured out. I'd suggest picking one or two of the following and trying to measure them in your own environment to understand whether these dynamics are showing up yet.
Human effort post-AI
If automation is introduced and handle time decreases on paper, but additional human effort is being created elsewhere that wasn't there before, the net picture may not match what your business case assumed. Measuring total human effort that follows an AI-attempted resolution reveals whether automation is removing work or redistributing it.
Demand origination mix
Categorising contacts as human-initiated, machine-initiated, or machine-initiated-then-escalated-to-human. We can't forecast what we can't categorise, and we can't match operating models to demand we haven't segmented.
Demand origination mix
The volume that AI "resolved" but which returns as a human contact within 48 hours. If your ghost demand rate is high, your deflection rate is overstating the effectiveness of automation.
Time-to-competent-response
How quickly can an agent match the knowledge level of an AI-informed customer? In an environment where customers arrive pre-armed, the time it takes for an agent to reach parity in the conversation is a meaningful indicator of whether your knowledge tools and augmentation are keeping pace.
These challenges are emerging rather than fully arrived. There are practical steps organisations can start taking that don't require a complete operational overhaul.
Rather than shrinking in the way many business cases predicted, contact centre demand is changing shape. Callers are changing capability, and existing operational models need to evolve alongside them. The organisations that start measuring and planning for these shifts now will be better positioned than those that wait until their existing metrics stop making sense.
Pick one of the four problems in this article, synchronised demand, the trust model, the fallback cascade, or the knowledge divide, and model what happens in your environment.
If you'd like help running that scenario or want to benchmark against what we're seeing across the industry, our workforce optimisation team can run a focused session with your planning and operations leads.
Customer expectations and behaviours are changing. Datacom helps organisations deliver better customer experiences through workforce optimisation, omnichannel support, real-time insights and scalable contact centre operations.
